With enterprise AI adoption failure rates as high as 95%, one expert suggests that companies are using the wrong playbook.
Siddharth Rao, a Global CIO with two decades of machine learning experience at firms like AstraZeneca, explains why the old rules of IT don't apply to AI.
He advises companies to ditch slow-moving pilot projects and embrace imperfect data to move from experimentation to execution.
Rao outlines a "two-bucket" strategy for building both defensive and offensive AI capabilities to avoid being left behind.
A strange contradiction sits at the heart of the current AI boom. While consumer tools like ChatGPT are adopted at a historic pace, enterprise-level success remains stubbornly elusive. Now, some reports indicate failure rates as high as 95%. In the search for a culprit, the blame often lands on budgets, data quality, or the technology itself. But the real issue is more fundamental: most companies are using the wrong playbook.
To understand why so many get it wrong, we spoke with Siddharth Rao, a global healthcare CIO and technology executive with a two-decade history in machine learning. Before ChatGPT became a household name, Rao was implementing machine learning at scale for companies like AstraZeneca, Biogen, and Charles River Laboratories. His experience in the heavily regulated healthcare sector offers a straightforward framework that other industries can replicate for success.
The disconnect stems from a misunderstanding of what makes AI different from prior technologies, Rao says. In the past, heavy lifting happened on the back end, with user experience focused on minimizing friction. But AI flips that script, empowering users on the front end. Essentially, the more you put in, the more you get out.
The pilot problem: The core problem is that companies fall back on the familiar but slow-moving playbook for typical IT projects, Rao says. They gather requirements, develop specifications, and launch small pilots. "The experimentation process can take several weeks or months, and you will lose leverage. Don't start with pilots. Pick one or two areas of your business where you see the maximum pain points and go full bore with an AI-first business strategy. Don't try to build guardrails before you even go in. That's going to slow you down."
Speed is everything right now, Rao says. As an example, he points to the healthcare industry, which is more than twice as far ahead in successful AI implementation as most other sectors. "AstraZeneca was doing this two decades ago. The big players were leveraging this technology at the grassroots level before anyone even knew what AI was because they realized the value of making that investment was exponential." Based on the healthcare playbook, Rao offers counterintuitive advice for any company looking to move from experimentation to execution.
Progress over perfection: "Don't let perfection be the enemy of good. Don't wait for a perfect data storage or data warehouse. Your data will never be perfect. Even if only a third of your data is usable, start with that." Still, many leaders are hesitant to use AI for documentation because its output isn't perfectly repeatable. Here, he offers OpenAI's own usage report as a counterargument. "If you look at the study data, two-thirds of the platform is still used to seek information. If you are not using this platform for building your training documentation, you're again missing out."
As companies mature, Rao suggests a "two-bucket" strategy. The first is a defensive play: an enterprise-wide tool that allows the company to monitor and govern the AI usage that is already happening. The second is the offensive play: a business-first strategy focused on "just-in-time solutioning" with custom models and deep integration.
Enter the agent army: For a glimpse of what this looks like at the extreme, Rao points to a cohort of Y Combinator startups. "Their entire customer support is autonomous. There's no human involved until it's a level-three or a level-four case. They have an army of 16 or 17 agents orchestrating things between each other, and there's just one conductor who's making sure that everything's going well. It's iterating through new features in a matter of hours, getting through testing in another matter of hours, and new features are making it to production literally every other day."
Now, that speed is translating to shocking growth, Rao concludes. With startups like no-code app builder Loveable scaling to over $75 million in annual recurring revenue in just seven months, the gap is only widening between AI-native companies and those still stuck in the pilot phase. "Just imagine how this is gonna scale up," Rao says. "It's really going to displace the way enterprise valuation metrics work and the way market shares have been historically defined."
